ActioDiff A Conceptual Framework for Modeling Intent Interaction and Prediction in Multi Agent Systems

Contemporary multi-agent systems generate vast amounts of interaction data, but existing formats fail to capture the semantic richness of agent intentions, decision-making processes, and predictive reasoning. While VectorDiff revolutionized dynamic data representation and SentioDiff enabled AI introspection, a critical gap remains in modeling the complex interplay between autonomous agents pursuing distinct goals through structured interactions. This paper introduces ActioDiff, a differential format that extends the baseScene + timeline paradigm to encompass multi-agent scenarios as first-class entities. ActioDiff captures not only what agents do, but why they act, how they interact, and what they predict will happen. Through its three-layer architecture - goalModel for intent representation, interaction protocols for structured communication, and hypothetical timelines for predictive reasoning - ActioDiff enables unprecedented transparency and analysis of multi-agent behavior. We demonstrate ActioDiff's effectiveness through a negotiation case study and discuss its implications for explainable AI, multi-agent coordination, and the analysis of emergent behavior.